Learning in multiagent systems

This dissertation investigates improving the performance of multiagent systems using machine learning strategies. As part of this study, three multiagent learning strategies were developed: the Learning Contract Net, the MAGE algorithm, and the Multiple Team Pursuit (MTP) algorithm.; The Learning Contract Net adds learning capabilities to Smith and Davis's Contract Net algorithm. This learning strategy reduces communication overhead by allowing communication to move from a broadcast to a direct addressing strategy. The result of this learning is to assign an organizational role to different agents from the perspective of contracting agents.; The MAGE algorithm learns information that models the other agents, their actions, and environmental states. It can synthesize models from two or more environmental states. The environment is used to learn the most likely actions of the agents and reduces the amount of communication required.; The MTP algorithm adds a new dimension to the well-known pursuit problem. The experimental results show cooperative problem-solving between multiple teams improves performance. The focus of these experiments was determining the optimum amount of cooperation in terms of how often the teams attempted to cooperate and for how long the teams waited to cooperate during these attempts. In addition, a domain-specific measure turned off cooperation attempts when a team had nearly completed its task. While this measure is domain specific, it showed that cooperation should be used only when there is a clear opportunity for a benefit.; General lessons learned include the following: simple types of cooperation can be effective; modeling agents and their actions is important, but the environment must also be modeled; credit assignment is difficult in single-agent learning systems, but it is even more difficult in multiagent systems; and organizational roles are important in a multiagent system since they help determine actions and responsibilities.